Overfitting

The most classic indicator: training loss continues to decrease monotonically while validation loss plateaus and then begins to increase. This inflection point marks where the model stops learning generalizable features and starts memorizing training set idiosyncrasies. In radiomics, this often manifests as the model latching onto scanner-specific noise rather than true tumor morphology.

• Monitor both curves on the same plot with a logarithmic y-axis
• The gap between curves quantifies the degree of overfitting
• Early stopping should trigger at the validation loss minimum

When a model's performance metric (e.g., AUC, C-index) varies dramatically across different cross-validation splits, it signals instability characteristic of overfitting. A well-generalized radiomic signature should yield consistent performance regardless of which patients are held out.

• Calculate the standard deviation of AUC across k-folds
• A coefficient of variation exceeding 10-15% warrants investigation
• Stratified splitting by clinical outcome and scanner type reduces misleading variance

When the number of selected radiomic features approaches or exceeds the number of clinical events (e.g., deaths, recurrences) in the training set, overfitting is mathematically guaranteed. This violates the rule of thumb requiring 10-20 events per candidate feature in Cox regression or logistic models.

• A model using 50 features on 30 events is certainly overfit
• LASSO regularization can help but does not eliminate the risk
• Internal bootstrap validation provides honest performance estimates

Achieving 100% accuracy or an AUC of 1.0 on training data is not a success—it is a red flag. In medical imaging, where biological variability and measurement noise are inherent, perfect discrimination indicates the model has discovered a data leakage pathway or memorized individual patient scans.

• Investigate for duplicate slices or patient overlap between splits
• Check if the model is exploiting scanner model or acquisition protocol as a confounder
• Realistic training AUC for radiomic tasks rarely exceeds 0.85-0.90

The definitive test: a radiomic model that performs well on internal hold-out data but collapses catastrophically on an independent external dataset from a different institution. This reveals that the model learned center-specific artifacts—such as reconstruction kernel, slice thickness, or contrast timing—rather than true tumor biology.

• External validation is the gold standard for proving generalizability
• Performance drops exceeding 15-20% AUC indicate severe overfitting
• ComBat harmonization can mitigate but not fully rescue an overfit signature

When adding more complex features (e.g., higher-order texture matrices, wavelet decompositions) yields diminishing or negative returns in validation performance, the model is fitting noise. A simple first-order histogram model that outperforms a complex GLCM+GLSZM+wavelet ensemble on validation data is a strong signal of overfitting in the complex model.

• Compare nested models with increasing complexity
• Use the Akaike Information Criterion (AIC) to penalize unnecessary parameters
• Parsimony is a virtue in clinically translatable radiomics

The fundamental tension between a model's ability to fit training data (low bias) and its stability across different training sets (low variance). Overfitting represents the high-variance extreme, where the model captures random fluctuations. The goal is the optimal tradeoff point that minimizes total error.

• High Bias: Underfitting, misses relevant relations
• High Variance: Overfitting, models noise
• Irreducible Error: Inherent noise in the data itself

A family of techniques that constrain model complexity to prevent overfitting by adding a penalty term to the loss function. L1 (Lasso) drives irrelevant feature weights to zero, performing feature selection. L2 (Ridge) shrinks all weights toward zero without eliminating them. Elastic Net combines both penalties.

• Dropout: Randomly deactivates neurons during training
• Early Stopping: Halts training when validation error rises
• Data Augmentation: Artificially expands training set diversity

A resampling procedure used to evaluate model generalization on limited data. k-Fold Cross-Validation partitions data into k subsets, training on k-1 folds and validating on the held-out fold, rotating until all data has served as validation. This provides a robust estimate of out-of-sample performance.

• Stratified k-Fold: Preserves class distribution in each fold
• Leave-One-Out (LOOCV): Extreme case where k equals sample size
• Nested CV: Outer loop for evaluation, inner loop for hyperparameter tuning

The expected prediction error on previously unseen data, decomposed into bias, variance, and irreducible noise. A model with low training error but high generalization error is overfit. The generalization gap—the difference between training and validation performance—is the primary diagnostic signal.

• Training Error: Performance on data the model has seen
• Validation Error: Proxy for true generalization error
• Test Error: Final unbiased estimate on held-out data

Least Absolute Shrinkage and Selection Operator—a linear regression method that performs both regularization and feature selection. By applying an L1 penalty proportional to the absolute value of coefficients, LASSO drives irrelevant feature weights exactly to zero, producing sparse models resistant to overfitting.

• Critical for high-dimensional radiomic datasets
• Selects a parsimonious set of predictive features
• Requires careful tuning of the lambda hyperparameter

A feature selection algorithm that identifies a subset of features with maximal statistical dependency on the target variable (relevance) and minimal mutual information among themselves (redundancy). By eliminating correlated features, mRMR reduces model complexity and overfitting risk in high-dimensional radiomic analyses.

• Mutual Information: Measures shared information between variables
• Greedy Search: Iteratively adds features optimizing the mRMR criterion
• Often paired with LASSO for robust radiomic signature construction